Doctoral Defense Announcement
Multi-Modal Intelligence for Multi-Hazard Response and Urban Resilience By
The capacity to transform multi-source natural hazard data into timely, trustworthy decisions has become increasingly critical as climate change intensifies both hazards and their interactions. This dissertation investigates how Generative AI and large language models (LLMs) can bridge three persistent challenges in disaster analytics: data heterogeneity, model interpretability and generalization, and practical applicability. The overarching objective is to establish a family of methods capable of integrating diverse data sources, leveraging generative and language-based reasoning, and producing actionable, explainable insights.
To achieve this goal, the dissertation explores multi-hazard response and resilience through the lens of multimodal intelligence, advancing the integration of causal and semantic reasoning in disaster response. The dissertation introduces a causality-informed framework for post-hurricane damage mapping that integrates InSAR and environmental indicators within a Bayesian network. By encoding hypothesized physical relationships, the model provides scalable assessments while revealing causal pathways from hazard drivers to structural outcomes, thereby enhancing both accuracy and explanatory depth. It further presents an LLM-driven rapid assessment pipeline for post-earthquake impact estimation, which continuously extracts, verifies, and synthesizes multilingual, crowdsourced reports to generate structured fatality evidence that provides timely updates to empirical assessment models, accelerating situation awareness during critical response periods. Another contribution advances a behavior-grounded LLM approach for wildfire evacuation modeling. By conditioning linguistic-based reasoning on constructs from behavioral theory, this method simulates and predicts evacuation intent that reflects cognitive, social, and contextual dimensions, improving interpretability and supporting transferability across communities and fire contexts. Extending beyond response, the dissertation also proposes a multi-agent generative design workflow for resilient streets and infrastructure. Coupling generative image models with evaluation agents and planning constraints, the framework generates and ranks alternative infrastructure designs that balance safety, accessibility, and aesthetics, operationalizing resilience by design through an iterative, human-in-the-loop process.
The dissertation establishes that coupling causal and semantic reasoning with multi-modal evidence enhances the speed, transparency, and generalizability of disaster intelligence, reinforcing the role of AI in emergency response and resilient infrastructure planning.